Cognitive diagnosis method based on hierarchical constraint perception

Abstract

The invention discloses a cognitive diagnosis method based on hierarchical constraint perception. The method comprises the following steps: S1, inputting student information to be processed; s2, performing hierarchical division on historical scores in the student information; s3, performing convolution feature extraction on the student information of the same level, and performing attention network aggregation to obtain knowledge features in the level; s4, aggregating student information of different levels, and forming a level knowledge proficiency degree vector by combining knowledge features in the levels and among the levels; s5, combining the hierarchical knowledge proficiency vector with the personalized knowledge proficiency vector to form a comprehensive proficiency vector, inputting the comprehensive proficiency vector into the cognitive diagnosis model, and predicting a practice question result; according to the method, cognitive diagnosis is carried out according to student information, an attention mechanism of internal and external knowledge states is introduced, the diagnosis precision is improved, the interpretability and reasoning ability of diagnosis are enhanced, and the result is more accurate and can be easily applied to teaching practice.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a cognitive diagnosis method based on hierarchical constraint perception and applied to educational assessment. Background Art

[0002] Cognitive diagnostics (CD) is a core task in educational data mining, which aims to reveal students' proficiency in specific knowledge concepts by analyzing their answer logs. With the popularization of intelligent education applications, the demand for cognitive diagnosis in assessing and improving individual development is growing. Although the existing cognitive diagnostic frameworks have improved the accuracy of diagnostic results to a certain extent by making full use of students' explicit response records, exercise texts, and exercise relationships, and combining considerations of special circumstances in the learning process, they mainly focus on the knowledge mastery level of individual students. Existing studies fail to fully reflect the relative level of students' ability performance in groups. Especially when dealing with group data, existing models often ignore the dynamic changes in ability differences between students, resulting in a lack of comprehensiveness in the inference of the ability level of each knowledge concept. This deficiency may limit the applicability of the evaluation results in actual educational environments. Therefore, accurately evaluating the relative level of students' performance in groups to ensure that the evaluation results are reasonable and meet the needs of the real educational environment is an issue that needs to be explored in depth. Solving this problem will promote the deeper development of the field of cognitive diagnosis and reflect the scientificity and fairness of modern educational evaluation. Summary of the invention

[0003] The embodiment of the present invention provides a cognitive diagnosis method based on hierarchical constraint perception to solve the above technical problems in the prior art and improve the accuracy and explainability of cognitive diagnosis in intelligent education.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] A cognitive diagnosis method based on hierarchical constraint perception includes the following steps:

[0006] Step S1: input the student information to be processed, the student information including the student ID, historical grades, exercise question information, and knowledge points involved in the questions;

[0007] Step S2: classify the historical scores in the student information into different levels to obtain the level information of the student;

[0008] Step S3: Perform convolution processing on the student information of the same level to extract features, aggregate the extracted features through the attention network, and obtain the knowledge features of the students in the level;

[0009] Step S4: randomly select a student from a different level, aggregate his information through the attention network, obtain the knowledge characteristics of students at different levels, and combine the knowledge characteristics of students within the level with the knowledge characteristics between levels to obtain the level knowledge proficiency vector;

[0010] Step S5: combining the student's hierarchical knowledge proficiency vector with the personalized knowledge proficiency vector to form the student's comprehensive proficiency vector, inputting the student's comprehensive proficiency vector and exercise question information into the cognitive diagnosis model to obtain the prediction result of the student's exercise question;

[0011] Furthermore, the specific method of step S1 is as follows:

[0012] S1.1 Obtain student information dataset;

[0013] S1.2 processes the data set, removes incomplete and invalid information (data is empty), and obtains valid student information, including: student ID information, historical grades, exercise question information, and knowledge points involved in the questions;

[0014] Furthermore, the specific method of step S2 is as follows:

[0015] S2.1: Calculate the mean and standard deviation of students' history scores. The calculation formula for the mean x and standard deviation σ of all students' history scores is:

[0016]

[0017] where x i represents the historical performance of the i-th student;

[0018] S2.2: Perform normal distribution based on the mean and standard deviation of students’ historical scores, complete student level mapping based on the distribution results, and obtain students’ level information. The specific formula is as follows:

[0019]

[0020] In the formula belongs to the hierarchical label set L={l0, l1, . . . , l g};

[0021] Furthermore, the specific method of step S3 is as follows:

[0022] S3.1: Convolution is performed on the student information at the same level to extract features, using different convolution kernels The extracted features are expressed as , the specific formula is as follows:

[0023]

[0024] Where C represents the convolution operation, is the knowledge point embedding matrix, is the feature dimension after convolution;

[0025] S3.2: Aggregate the extracted features through the attention network to get the output of the attention model and calculate the query for each convolution head h ,key Sum The specific formula is:

[0026]

[0027]

[0028]

[0029] in is a trainable weight matrix for queries, keys, and values, which performs similarity measurement based on the feature representation generated by convolution and generates attention output , the specific formula is as follows:

[0030]

[0031] in is the Sigmoid function;

[0032] S3.3: Aggregate the attention outputs of all convolution heads to obtain ,Will Further transformation to obtain the knowledge characteristics of students within the level , the specific formula is as follows:

[0033]

[0034]

[0035] in is the number of attention heads, is the weight matrix of the fully connected layer, is the bias term.

[0036] Furthermore, the specific method of step S4 is as follows:

[0037] S4.1: Randomly select a student from different levels and aggregate their information through the attention network. The formula is as follows:

[0038]

[0039] in Indicates the total number of other levels, g indicates the current level, Indicates other levels, Context-aware information for different levels of student information aggregated through the attention network;

[0040] S4.2: Combine the knowledge characteristics of students within the level and the knowledge characteristics between levels to obtain the level knowledge proficiency vector, the formula is:

[0041]

[0042] Where |H| is the number of levels;

[0043] Furthermore, the specific method of step S5 is as follows:

[0044] S5.1: The student's hierarchical knowledge proficiency vector and personalized knowledge proficiency vector are combined using an adaptively optimized personalized weighting strategy to form the student's comprehensive proficiency vector. The calculation formula is as follows:

[0045] Students' personalized knowledge proficiency vector The knowledge point relevance vector of the exercise questions The calculation formula is:

[0046]

[0047]

[0048] Where S∈R n×k is a trainable knowledge proficiency matrix, k is the number of knowledge points, and the Q matrix is ​​recorded as:

[0049]

[0050] In this matrix, if exercise e i Related knowledge points kc j , then Q ij = 1; otherwise, it is Q ij = 0;

[0051] The student's overall proficiency vector , the specific formula is as follows:

[0052]

[0053]

[0054] in, is the trainable weight matrix, s is the student index;

[0055] S5.2: Input the student's comprehensive proficiency vector and exercise question information into the cognitive diagnosis model to obtain the prediction results of the student's exercise questions. The specific formula is as follows:

[0056]

[0057] in Student responses to exercises predict outcomes, is the knowledge point relevance vector of the exercise question, and CDM() is the cognitive diagnosis model;

[0058] In combination with step S1 to step S5, the student's comprehensive proficiency vector and exercise question information are input into the cognitive diagnosis model to obtain the prediction results of the student's exercise questions, and further include: using a cross entropy loss function to measure the difference between the predicted value y and the student's actual answer label r. The specific loss function is as follows:

[0059]

[0060] Beneficial effects:

[0061] This method proposes a cognitive diagnosis method based on hierarchical constraint perception. This method effectively limits the changes in students' knowledge status, making it more reasonable and more in line with the needs of real educational scenarios. It introduces an attention mechanism for knowledge status inside and outside the hierarchy, significantly improving the diagnostic accuracy of hierarchical constraint perception, thereby more accurately identifying and evaluating students' cognitive status. Through the above mechanism, the interpretability and reasoning ability of cognitive diagnosis are improved, making the diagnostic results not only accurate but also easy to understand and apply to teaching practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0063] Figure 1 This is a flow chart of a cognitive diagnosis method based on hierarchical constraint perception of the present invention;

[0064] Figure 2 Density kernel plots for hierarchical constraints on PISA-Reading and PISA-Mathematics datasets;

[0065] Figure 3 It is a graph showing the knowledge mastery of 11 knowledge points by students at six different levels;

[0066] Figure 4This is the fitting diagram of personalized ability and hierarchical ability after averaging the ability of each knowledge point of 30 students in the PISA-Science data set;

[0067] Figure 5 Personalization capabilities , level capabilities and comprehensive capabilities Radar chart of the relationship between the difficulty of the question and the difficulty of the question. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] See also Figure 1 This is a flow chart of a cognitive diagnosis method based on hierarchical constraint perception provided by an embodiment of the present application, and the method includes the following steps:

[0070] Step S1: input the student information to be processed, the student information including the student ID, historical grades, exercise question information, and knowledge points involved in the questions;

[0071] Step S2: classify the historical scores in the student information into different levels to obtain the level information of the student;

[0072] Step S3: Perform convolution processing on the student information of the same level to extract features, aggregate the extracted features through the attention network, and obtain the knowledge features of the students in the level;

[0073] Step S4: randomly select a student from a different level, aggregate his information through the attention network, obtain the knowledge characteristics of students at different levels, and combine the knowledge characteristics of students within the level with the knowledge characteristics between levels to obtain the level knowledge proficiency vector;

[0074] Step S5: combining the student's hierarchical knowledge proficiency vector with the personalized knowledge proficiency vector to form the student's comprehensive proficiency vector, inputting the student's comprehensive proficiency vector and exercise question information into the cognitive diagnosis model to obtain the prediction result of the student's exercise question;

[0075] It is worth mentioning that when conducting cognitive diagnosis on students, the device used is a 64-bit Ubuntu 20.04.5 LTS server equipped with a 2.3GHz InterXeon Gold 5218 CPU and a 32GB Tesla V100 GPU, and the calculation process is based on the PyTorch framework.

[0076] Table 1

[0077]

[0078] It is worth noting that the input data of this example uses the 2015 International Student Assessment Program (PISA 2015) public dataset. This dataset covers more than 500,000 students from 73 countries and regions. Three sub-datasets of science, reading and mathematics were extracted from it. Questions marked as "Full credit" in the dataset are considered to be answered correctly (labeled as 1), while "No credit" and "Partial credit" are considered to be answered incorrectly (labeled as 0), and in other cases, the student is considered to have not answered the question. To ensure the reliability of the analysis results, student data with less than 30 records were removed, and student data with less than 20 records were removed from the mathematics dataset. Table 1 shows the basic statistical information of these datasets. Finally, all datasets are randomly divided into 70% for training, 10% for validation, and 20% for testing to ensure the effectiveness of the method evaluation.

[0079] The student ID and exercise question information in the data set are encoded to generate personalized knowledge proficiency vectors and knowledge point relevance vectors of the exercise questions. ∈{0, 1} 1×n , where n represents the number of students, and the exercise information is encoded in a one-hot format ∈{0, 1} 1×m Where m represents the number of practice questions, and the student's personalized knowledge proficiency vector The knowledge point relevance vector of the exercise questions The calculation formula is:

[0080]

[0081]

[0082] Where S∈R n×k is a trainable knowledge proficiency matrix, k is the number of knowledge points, and the Q matrix is ​​recorded as:

[0083]

[0084] In this matrix, if exercise e i Related knowledge points kc j , then Q ij = 1; otherwise, it is Q ij = 0.

[0085] Divide the historical scores in the student information into levels, obtain the level information of the student, perform one-hot encoding on the student level information and convert it into a trainable proficiency vector. Let X={x1,x2, . . . ,x n} is the historical scores of all students, where x i represents the history score of the i-th student. The calculation formula for the mean x and standard deviation σ of all students' history scores is:

[0086]

[0087]

[0088] Based on this, we define the score interval B={b0,b1, . . . , b g} and its corresponding hierarchical label set L={l0, l1,. . . , l g}. The mapping function f:X→L is defined as:

[0089]

[0090] Where b j ≤x i j+1 ,∀i∈[1, n],j∈[0, g], embed the student-level entries through one-hot encoding, and use O∈R n×g Indicates, specifically:

[0091]

[0092] In the formula, ∀i∈[1, n],j∈[0, g-1], the O vector is converted into a learnable knowledge point vector H through linear transformation KCs ∈R n×k , specifically expressed as:

[0093]

[0094] In the formula ∈R g×k is the weight matrix of the linear layer, b∈R k is the bias vector, g is the number of levels, and k is the number of knowledge points.

[0095] ​The trainable proficiency vectors of students at the same level are convolved to extract features, and the extracted features are aggregated through the attention network to obtain the knowledge features of students in the same level through different convolution kernels. The extracted features are expressed as The specific formula is as follows:

[0096]

[0097] Where C represents the convolution operation, is the knowledge point embedding matrix, is the feature dimension after convolution, and the query is calculated for each convolution head h ,key Sum The specific formula is:

[0098]

[0099]

[0100]

[0101] in is a trainable weight matrix for queries, keys, and values, which performs similarity measurement based on the feature representation generated by convolution and generates attention output , the specific formula is as follows:

[0102]

[0103] in is the Sigmoid function, which is used to emphasize the interdependence between them. Finally, the attention outputs of all convolution heads are aggregated to obtain In order to further enhance the expressive power of feature representation and help the model better capture the complex relationship between different knowledge points, Further transformation to obtain the knowledge characteristics of students within the level , the specific formula is as follows:

[0104]

[0105]

[0106] in is the number of attention heads, is the weight matrix of the fully connected layer, is the bias term.

[0107] Randomly select a student from different levels, aggregate their trainable proficiency vectors through the attention network, obtain the knowledge features between different levels of students, and combine the knowledge features of students within the level with the knowledge features between levels to obtain the level knowledge proficiency vector. In order to reduce the computational complexity, randomly select a student feature from each level from other levels for interaction. This design not only helps to reduce the amount of calculation, but also ensures that student information from different levels is obtained, thereby enhancing the generalization ability of the model. The specific definition is as follows:

[0108]

[0109]

[0110]

[0111] Where g represents the current level, To represent other levels, in order to highlight the correlation of knowledge between levels and make the student ability assessment more sensitive, the attention score between the query at level g and the key of the sampled level g′ student subset is calculated to obtain the context-aware representation formula as follows:

[0112]

[0113] This process enhances the comprehensiveness and accuracy of the student ability assessment by fusing information from different levels, and then averages the sampled context representations from multiple other levels g′ to integrate the context information from all other levels, as follows:

[0114]

[0115] in Represents the total number of other levels and is used to calculate the average value, which can effectively integrate knowledge information from multiple levels, thereby further improving the robustness of the model and the ability to evaluate the student's knowledge status. Finally, the output of the intra-level attention is combined with the sampling-based inter-group context information to obtain the final representation of the hierarchical knowledge proficiency:

[0116]

[0117] Here, |H| is the number of levels. This combination not only utilizes the contextual information within the level, but also introduces the interaction between levels, thus ensuring that the model is more comprehensive and accurate in estimating student abilities.

[0118] The student's hierarchical knowledge proficiency vector is combined with the personalized knowledge proficiency vector to form the student's comprehensive proficiency vector. The student's comprehensive proficiency vector and the unique hot encoding of the exercise are input into the cognitive diagnosis model, and the student's prediction result for the exercise is obtained. Describe the individual characteristics of each student, use adaptive optimization and personalized weighting strategy to classify the level of proficiency Combined with personalized features , forming the student's comprehensive proficiency vector , the specific formula is as follows:

[0119]

[0120]

[0121] in, is a trainable weight matrix, s is the student index, and the basic paradigm of the cognitive diagnosis model is used to predict the student's response to the exercise. It is applicable to a variety of cognitive diagnosis models, such as item response theory (IRT), multidimensional item response theory (MIRT), deterministic inputs, noisy "and" model (DINA), and neurocognitive diagnosis model (NCDM). The prediction result of the student's response to the exercise is , the specific formula is as follows:

[0122]

[0123] in is the knowledge point relevance vector of the exercise questions, and CDM() is the cognitive diagnosis model.

[0124] The student's comprehensive proficiency vector and the unique-hot encoding of the exercise are input into the cognitive diagnosis model. After obtaining the student's prediction results for the exercise, the following steps are also included: using the cross entropy loss function to measure the difference between the predicted value y and the student's actual answer label r. The specific loss function is as follows:

[0125]

[0126] Table 2

[0127]

[0128] See Table 2, which shows the prediction effect data of this method on three PISA data sets. Four implementation plans based on this method are proposed. The cognitive diagnosis method based on hierarchical constraint perception (Hierarchy Constraint-Aware Cognitive Diagnosis, referred to as: HCD) is combined with typical cognitive diagnosis models, namely HCD-IRT, HCD-MIRT, HCD-DINA and HCD-NCDM. Through fine parameter tuning, it is ensured that each model runs in the best state, thereby ensuring the fairness of the comparison results. The area under the curve (AUC), accuracy (ACC) and root mean square error (RMSE) are used as the main performance indicators to comprehensively measure the effectiveness of each model. Table 2 shows the experimental results. Through the analysis of Table 2, several key findings are obtained. First, this method can greatly improve the prediction accuracy of traditional cognitive models. Second, after applying the NCDM model to this method, the optimal performance is achieved in terms of AUC, ACC and RMSE indicators. Third, after analyzing the number of three data sets, it is found that PISA-Science has the largest amount of data, PISA-Reading has the least amount of data, and PISA-Mathematics has the least interactive data. On these three data sets, the indicators reflected after embedding this method are better than typical cognitive models (IRT, MIRT, DINA, NCDM), proving that this method has good adaptability. Fourth, the two key steps of this method, step S3 (CEA) and step S4 (RSA), have made great contributions to the final performance of this method. The data in Table 2 are verified after removing these two steps. It is found that removing any of these two steps will have a significant impact on the performance, especially after removing step S3, the performance is seriously degraded.

[0129] See also Figure 2 , Figure 2 The density kernel diagram is used to show the performance of hierarchical constraints on two typical data sets, PISA-Reading and PISA-Mathematics. Figure 2It was found that, first, overall, both NCDM and the HCD-NCDM model embedded in this method can distinguish the students' ability distribution well on the two data sets, and the performance trend is consistent with the average level of the students' prior distribution, which shows that the existing cognitive diagnosis technology is effective in assessing students' knowledge status. Secondly, by comparing the performance of the two data sets, it is found that on the PISA-reading data set with a small number of students, although the average level of knowledge status learned by the two models is similar to the students' prior distribution, the overall distribution state is far from the normal distribution; while on the PISA-mathematics data set with a large amount of data, the students' ability distribution is closer to the normal distribution, which is also in line with the general law of human ability development. Third, the students' knowledge status learned by the NCDM model on the PISA-mathematics data set fluctuates greatly, while after being embedded in this method, the students' ability performance is smoother and more consistent with the students' prior statistical distribution. This shows that the students' abilities learned by the cognitive diagnosis model embedded in this method are not only more reasonable, but also can avoid ignoring the students' relative ability performance just to improve the model accuracy, thus having better interpretability. Finally, through further observation of the two data sets, especially the distribution of students with an ability of 0.5, it is found that the distribution density of students in the model embedded in this method at this stage is consistent with the prior distribution. For example, in the PISA-Mathematics data set, the proportion of students with an ability of 0.5 in the prior distribution is about 24%, while the NCDM model only learns 13.5%; on the contrary, the HCD-NCDM model has a student proportion of 28% at this stage, which is closer to the prior distribution. This once again shows that this method has a positive effect on the cognitive diagnosis model to more reasonably reflect the knowledge status of students.

[0130] See also Figure 3 , Figure 3 It shows the influence of hierarchical constraints on individuals. Students are divided into six levels, and a student is randomly selected from each level to intuitively show their mastery of 11 knowledge points in the PISA-Mathematics dataset. Figure 3 It was found that, first, the knowledge status of students at different levels showed significant differences. Since each student's answer record was different, the knowledge mastery was diverse. This shows that this method has a strong discernment in capturing students' knowledge status. Second, regardless of the level, the mastery of some knowledge points is generally low. For example, at the H1 level, the mastery of multiple knowledge points is close to 0.24. This shows that it is because the students did not involve the relevant knowledge points. This method can not only distinguish the knowledge mastery of different students, but also reveal the common learning weaknesses, providing teaching staff with a powerful tool to optimize teaching methods and content, thus playing an important role in the field of smart education.

[0131] See also Figure 4 , Figure 4 This is a fitting diagram of the personalized ability and hierarchical ability of 30 students in the PISA-Science data set after averaging the ability of each knowledge point. Figure 4 Discover, first of all, the individual abilities of each student and hierarchical capabilities The results show significant differences, which illustrates the importance of personalized student ability modeling. Secondly, the weights of personalized ability and hierarchical ability on students are different. For example, for student 1, his personalized ability level is better than the hierarchical ability level, but the hierarchical ability plays a greater role in the final student ability θ. On the contrary, for student 14, although the hierarchical ability level is higher, the influence of personalized ability on the overall ability is greater. This shows that each student's ability is affected differently by the two characteristics. Finally, we note that the slope of the regression line for hierarchical ability is higher than that for personalized ability, indicating that in the performance of the current 30 students, the overall impact of hierarchical ability on the final student ability is greater than the impact of personalized ability.

[0132] See also Figure 5 , Figure 5 Personalization capabilities , level capabilities and comprehensive capabilities The relationship between the student's cognitive diagnosis results and his / her ability and the difficulty of the questions can be quantitatively analyzed using formula (24). A student in the PISA-Science dataset and the five questions he / she answered were randomly selected. Figure 5 The left half shows the student's answer, and the right half shows the model's diagnosis results. The radar chart shows the student's mastery of each knowledge point and the difficulty of the question. It can be seen from the figure that when the student's knowledge mastery meets the requirements of the question, they are more likely to give the correct answer. For example, for question e3, the student's personalized ability and hierarchical capabilities are 0.64 and 0.60 respectively, the comprehensive ability θ = 0.63, and the difficulty of the question is 0.41. Since the student's proficiency in the knowledge points involved in the question is higher than the required level, he successfully answered the question. On the contrary, in question e1, the student's personalized ability and hierarchical capabilities are 0.795 and 0.37 respectively, the comprehensive ability θ = 0.60, and the difficulty of the question is 0.793. Although the student's individual ability is higher than the difficulty of the question, his overall ability is lower than the ability required to solve the question, and he ultimately failed to answer correctly. This phenomenon is consistent with the student's response, which once again shows that this method can significantly improve the accuracy of cognitive diagnosis.

Claims

1. A cognitive diagnosis method based on hierarchical constraint perception, characterized in that: The following steps are involved: Step S1: input the student information to be processed, the student information including the student ID, historical grades, exercise question information, and knowledge points involved in the questions; Step S2: classify the historical scores in the student information into different levels to obtain the level information of the student; Step S3: Perform convolution processing on the student information of the same level to extract features, aggregate the extracted features through the attention network, and obtain the knowledge features of the students in the level; Step S4: randomly select a student from a different level, aggregate his information through the attention network, obtain the knowledge characteristics of students at different levels, and combine the knowledge characteristics of students within the level with the knowledge characteristics between levels to obtain the level knowledge proficiency vector; Step S5: Combine the student's hierarchical knowledge proficiency vector with the personalized knowledge proficiency vector to form the student's comprehensive proficiency vector, input the student's comprehensive proficiency vector and exercise question information into the cognitive diagnosis model to obtain the prediction results of the student's exercise questions.

2. A cognitive diagnosis method based on hierarchical constraint perception according to claim 1, characterized in that: The specific method of step S1 is as follows: S1.1 Obtain student information dataset; S1.2 processes the data set, removes incomplete and invalid information (data is empty), and obtains valid student information, including: student ID information, historical grades, exercise question information, and knowledge points involved in the questions.

3. The cognitive diagnosis method based on hierarchical constraint perception according to claim 1 is characterized in that: The specific method of step S2 is as follows: S2.1: Calculate the mean and standard deviation of students' history scores. The calculation formula for the mean x and standard deviation σ of all students' history scores is: where x i represents the historical performance of the i-th student; S2.2: Perform normal distribution based on the mean and standard deviation of students’ historical scores, complete student level mapping based on the distribution results, and obtain students’ level information. The specific formula is as follows: In the formula belongs to the hierarchical label set L={l0, l1, . . . , l g }.

4. The cognitive diagnosis method based on hierarchical constraint perception according to claim 1 is characterized in that: The specific method of step S3 is as follows: S3.1: Convolution is performed on the student information at the same level to extract features, using different convolution kernels The extracted features are expressed as , the specific formula is as follows: Where C represents the convolution operation, is the knowledge point embedding matrix, is the feature dimension after convolution; S3.2: Aggregate the extracted features through the attention network to get the output of the attention model and calculate the query for each convolution head h ,key Sum The specific formula is: in is a trainable weight matrix for queries, keys, and values, which performs similarity measurement based on the feature representation generated by convolution and generates attention output , the specific formula is as follows: in is the Sigmoid function; S3.3: Aggregate the attention outputs of all convolution heads to obtain ,Will Further transformation to obtain the knowledge characteristics of students within the level , the specific formula is as follows: in is the number of attention heads, is the weight matrix of the fully connected layer, is the bias term.

5. The cognitive diagnosis method based on hierarchical constraint perception according to claim 1, characterized in that: The specific method of step S4 is as follows: S4.1: Randomly select a student from different levels and aggregate their information through the attention network. The formula is as follows: in Indicates the total number of other levels, g indicates the current level, Indicates other levels, Context-aware information for different levels of student information aggregated through the attention network; S4.2: Combine the knowledge characteristics of students within the level and the knowledge characteristics between levels to obtain the level knowledge proficiency vector, the formula is: Where |H| is the number of levels.

6. The cognitive diagnosis method based on hierarchical constraint perception according to claim 1, characterized in that: The specific method of step S5 is as follows: S5.1: The student's hierarchical knowledge proficiency vector and personalized knowledge proficiency vector are combined using an adaptively optimized personalized weighting strategy to form the student's comprehensive proficiency vector. The calculation formula is as follows: Students' personalized knowledge proficiency vector The knowledge point relevance vector of the exercise questions The calculation formula is: Where S∈R n×k is a trainable knowledge proficiency matrix, k is the number of knowledge points, and the Q matrix is ​​recorded as: In this matrix, if exercise e i Related knowledge points kc j , then Q ij = 1; otherwise, it is Q ij = 0; The student's overall proficiency vector , the specific formula is as follows: in, is the trainable weight matrix, s is the student index; S5.2: Input the student's comprehensive proficiency vector and exercise question information into the cognitive diagnosis model to obtain the prediction results of the student's exercise questions. The specific formula is as follows: in Student responses to exercises predict outcomes, is the knowledge point relevance vector of the exercise questions, and CDM() is the cognitive diagnosis model.

7. The method according to any one of claims 1 to 6, characterized in that After inputting the student's comprehensive proficiency vector and exercise question information into the cognitive diagnosis model and obtaining the predicted results of the student's exercise questions, it also includes: using the cross entropy loss function to measure the difference between the predicted value y and the student's actual answer label r. The specific loss function is as follows:

Images